Aethex checks an automated decision against hard rules and returns a verdict — approve, escalate, or block — with the exact rule that fired. No AI model sits in the decision path, so the verdict can't drift, and the same input always returns the same answer.
Run it a thousand times — the hash never changes. That is the entire point.
AI can generate a decision. It cannot be trusted to execute one.
Every AI agent in production has the same gap: the model decides, and the system acts. There is nothing in between to ask should this actually happen — and can we defend it later?
The common answer is to use more AI to watch the AI. But a probabilistic system cannot audit another probabilistic system and call the result defensible — the watcher can drift and hallucinate too.
Aethex is the missing layer: a rule-based verdict on every decision, with the reasoning shown, so a person, an auditor, or a regulator can see exactly why.
The same ten-engine core runs across every domain through a single call. Only the rule vocabulary changes.
Sanctions screening (OFAC / UN / EU), AML typologies — structuring, layering, cuckoo smurfing — SWIFT anomalies, KYC/CDD and PEP exposure. Each verdict cites the regulation and the rule that fired.
LivePost-compromise attack-chain analysis mapped to MITRE ATT&CK. Answers the SOC's real question — isolate, escalate, or ignore — and keeps ordinary noise at LOW instead of drowning analysts in false positives.
LiveEvery verdict ships with the rule, the input that satisfied it, a full engine-by-engine breakdown, and a disagreement score. Built to be read by a regulator or an examiner, not just a dashboard.
LiveVerification for actions proposed by AI agents — destructive commands, invariant violations, sensitive-system access — checked against stated rules before they execute.
In developmentAccess by REST API, an interactive sandbox, or the Python SDK — pip install aethex
Each engine implements a distinct formal reasoning system. None of them is an LLM. Their agreement — or disagreement — is made explicit and measurable.
Typed events plus a domain. Customer vocabulary is resolved to Aethex's canonical rules. Below 50% recognition, Aethex refuses to guess and says so.
Pattern, grammar, causal logic, threat-depth, compression, temporal and stakeholder context — each reasons independently over the same events.
Determinate violations decide instantly; weak signals need a second source to escalate. A single alarmed engine is a signal, not a verdict.
Severity, confidence, the rule that fired, the recommended action with its citation, and the reasoning trail. Reproducible on demand.
Model-based tools produce a risk score and then generate a justification after the fact. When the score is wrong, so is the story. Aethex makes the decision with rules, so the explanation isn't generated — it is the decision.
The one person clearing the alert queue and signing the reports at a money-moving fintech or exchange. Aethex turns a defensible answer from a scramble into a record that already exists.
Analysts buried under thousands of daily alerts, most of them false. Aethex returns rule-backed verdicts so the alerts that reach a human are the ones worth their time.
Anyone putting an agent in a position to take real actions who needs a deterministic check between what the agent intends and what actually executes.
Compose a set of events, pick a domain, and watch the engines return a verdict with the reasoning shown. No signup.